Deep learning framework leveraging graph-based retinal layers segmentation in volumetric OCT with custom CNN for multiple sclerosis detection

Multiple sclerosis (MS) is a demyelinating disorder of the central nervous system where retinal layer thinning, measured via optical coherence tomography (OCT), has emerged as a promising biomarker. However, the development of automated, explainable, and efficient computational frameworks for OCT-based MS detection remains an unmet need. In this study, we propose a novel multi-stage graph-based automatic segmentation approach to delineate seven different retinal layers in macular-centered OCT B-scans. From the segmented layers, thickness profiles were extracted to feed into custom convolutional neural network (CNN) for MS vs. healthy control classification. Proposed custom CNN achieved an accuracy of 94%, sensitivity of 92%, and specificity of 97%, outperforming several pre-trained established architectures (e.g., VGG16, ResNet). Statistical analysis (one-way ANOVA) revealed significant thinning in six out of seven retinal layers in MS patients compared to controls (p<0.05). Grad-CAM explanation confirmed that the model focused on anatomically feasible regions, particularly the ILM and NFL-GCL complex. Integration of novel automatic segmentation, validated thickness biomarkers, light-weight custom CNN establishes an explainable, automated pipeline for OCT B-scan based MS detection. Proposed framework demonstrates a strong potential for clinical translation as a non-invasive, imaging-based diagnostic aid for MS detection.